Ultra-Processed Food: A Consistent Signal, a Contested Category

The phrase has escaped the journals. Shoppers now turn boxes over in the aisle looking for something they have been told to avoid, without much agreement on what they are looking for. Some scan for the number of ingredients. Some scan for anything unpronounceable. Some have decided that a packet is the problem and a jar is not. None of those is the published rule, and the gap between the popular version of the term and the technical one is where a good deal of the confusion — and a fair amount of unnecessary anxiety — lives.

What The NOVA Rule Actually Says

The term comes from NOVA, a classification set out by Carlos Monteiro and colleagues. The 2019 paper most often cited for it is worth being precise about: it is a commentary in Public Health Nutrition, not a study. It has no participants, no duration and no outcomes. It describes itself as containing “a clear and simple guide designed to identify ultra-processed foods” Monteiro 2019. What it offers is a rule, and the rule is narrower than the aisle version of the term.

NOVA classifies foods “according to the extent and purpose of the industrial processing they undergo”, considering “all physical, biological and chemical methods used during the food manufacturing process, including the use of additives” Monteiro 2019. The practical test the guide gives for group four is whether a product’s ingredient list contains at least one item characteristic of that group: either a food substance of no or rare culinary use, or one of a set of additive classes the paper calls cosmetic. The substances it names are varieties of sugars (fructose, high-fructose corn syrup, fruit juice concentrates, invert sugar, maltodextrin, dextrose, lactose), modified oils (hydrogenated or interesterified) and protein sources including hydrolysed proteins, soya protein isolate, gluten, casein, whey protein and mechanically separated meat. The cosmetic additive classes it names are flavours, flavour enhancers, colours, emulsifiers, emulsifying salts, sweeteners, thickeners, and anti-foaming, bulking, carbonating, foaming, gelling and glazing agents Monteiro 2019. As the rule is written, the test turns on which of those items appears on the list, not on how long the list runs.

The guide is also explicit about what the manufacturing is for. Fractioning whole foods into substances, chemically modifying those substances, reassembling modified and unmodified substances, frequent use of cosmetic additives and sophisticated packaging — these processes are described as designed to create products that are highly profitable, convenient and hyper-palatable, and “liable to displace all other NOVA food groups” Monteiro 2019. It is worth being clear that the authors do not present their category as a neutral description of factory technique. Their stated position, in the same paper, is that ultra-processing “makes them highly profitable, intensely appealing and intrinsically unhealthy”, while they simultaneously argue that “it is therefore unhelpful to criticise foods as being ‘processed’” at all, since almost all food is processed to some degree, and that diets restricted to unprocessed food “would be less diverse and less secure” Monteiro 2019. The same group set out that thesis and the evidence they say supports it at length in a 2025 Lancet Series paper Monteiro 2025.

The Four NOVA Groups, As The Guide Defines Them

Group one is unprocessed or minimally processed foods; the guide’s worked examples include plain steel-cut oats, plain corn flakes and shredded wheat. Group two is processed culinary ingredients — “substances obtained directly from group 1 foods or from nature, like oils and fats, sugar and salt”, created by pressing, centrifuging, refining, extracting or mining, and used in preparing, seasoning and cooking group one foods. Group three is processed foods: “industrial products made by adding salt, sugar or other substance found in group 2 to group 1 foods, using preservation methods such as canning and bottling, and, in the case of breads and cheeses, using non-alcoholic fermentation” Monteiro 2019.

Group four is ultra-processed: “formulations of ingredients, mostly of exclusive industrial use”. The paper’s own example list runs through carbonated soft drinks; sweet or savoury packaged snacks; chocolate and confectionery; ice cream; mass-produced packaged breads and buns; margarines and other spreads; biscuits, pastries, cakes and cake mixes; breakfast “cereals”; pre-prepared pies, pasta and pizza dishes; poultry and fish nuggets and sticks, sausages, burgers, hot dogs and other reconstituted meat products; and powdered or packaged instant soups, noodles and desserts Monteiro 2019.

Two of the guide’s worked examples do more explanatory work than the list does. On bread: “Industrial breads made only from wheat flour, water, salt and yeast are processed foods, while those whose lists of ingredients also include emulsifiers or colours are ultra-processed.” On cereal: “Plain steel-cut oats, plain corn flakes and shredded wheat are minimally processed foods, while the same foods are processed when they also contain sugar, and ultra-processed if they also contain flavours or colours” Monteiro 2019. Read literally, added sugar alone does not move a cereal into group four under the guide’s own worked example; a flavour or a colour does. That sits awkwardly beside the same paper’s example list, which names breakfast “cereals” among group four outright. The tension is inside the source, and it is a fair warning that the boundary is finer than the shopping-aisle version of the term suggests.

The Randomised Feeding Trials

Two randomised crossover feeding trials carry most of the weight here, and a third, much smaller crossover study sits alongside them; they were built to answer slightly different questions.

The first was run inside a metabolic ward. Twenty weight-stable adults, mean age 31.2 years and mean body mass index 27 kg/m², were randomised to an ultra-processed or an unprocessed diet for two weeks, then immediately switched to the other for two weeks — 28 days in total, with no weight-maintenance run-in and no washout between diets. Meals were designed to be matched for presented calories, energy density, macronutrients, sugar, sodium and fibre, and participants were instructed to eat as much or as little as they wanted. Energy intake was 508 ± 106 kcal a day greater on the ultra-processed diet (P=0.0001), meal eating rate was greater (17 ± 1 kcal/min, P<0.0001), and participants gained 0.9 ± 0.3 kg on the ultra-processed diet (P=0.009) while losing 0.9 ± 0.3 kg on the unprocessed one (P=0.007) Hall 2019.

The trial’s null results matter nearly as much as its positive ones. Once adjusted for energy intake, reported hunger, fullness, satisfaction and capacity to eat did not differ significantly between the diets, and neither did the rated pleasantness or familiarity of the meals. Total fibre and total sugar intakes were not significantly different either, although sodium intake was higher on the ultra-processed diet (5.8 ± 0.2 versus 4.6 ± 0.2 g a day, P<0.0001) Hall 2019. The authors also flag that the matched energy density was achieved partly by including beverages that carried dissolved fibre supplements; the non-beverage foods were about 85 per cent more energy dense on the ultra-processed side (2.147 versus 1.151 kcal/g), which they say “likely contributed to the observed excess energy intake”. They state the limit of their own design plainly: “Our study was not designed to identify the cause of the observed differences in energy intake.” They add that the metabolic-ward setting makes the results difficult to generalise to free-living conditions, and that a larger difference might have appeared had the two menus differed in sugar, fat and sodium as much as real-world diets typically do Hall 2019. Twenty participants is a small sample.

The missing washout in that design has since been examined directly. A 2024 analysis reported that diet order significantly affected energy balance in crossover studies without a washout period for diets varying in macronutrients, but not for diets varying in ultraprocessing Sciarrillo 2024. A separate randomised, open-label crossover study published the same year is the smallest piece of randomised evidence on this page, and it should be read as such: it gave nine overweight or obese Japanese men one week of ultra-processed food and one week without, in random order, separated by a two-week washout. Participants gained 1.1 kg more weight during the ultra-processed week (95% CI 0.2 to 2.0; P=0.021), consumed 813.5 kcal more a day (342.4 to 1284.7; P=0.0041), and recorded significantly fewer chews per calorie (P=0.016) Hamano 2024. Nine single-sex participants over one week can carry very little weight on their own: that sample says nothing about women, and a week is far too short to speak to the trajectory of body weight over a season, let alone a lifetime.

The free-living crossover trial went after the objection that inpatient work happens in a setting nothing like anybody’s kitchen. Fifty-five adults in England, all with a body mass index between 25 and 40 kg/m² and a habitual ultra-processed intake of at least half their daily calories, were provided with two eat-as-much-as-you-like diets in random order, eight weeks on each, delivered to their homes. Both diets were built to follow the UK’s Eatwell Guide; fifty participants made up the intention-to-treat sample. Both arms produced weight loss, and the minimally processed arm produced more of it: −2.06 per cent of body weight (95% CI −2.99 to −1.13) against −1.05 per cent (−1.98 to −0.13), a difference of −1.01 percentage points (−1.87 to −0.14, P=0.024; Cohen’s d −0.48). Mild gastrointestinal adverse events were common on both diets Dicken 2025.

Two features of that trial are easy to miss and change how the result should be read. Its ultra-processed arm was deliberately a nutritionally improved, reformulated one — the authors describe including reformulated breakfast cereals, ready meals and plant-based alternatives so that the ultra-processed diet still delivered recommended intakes of nutrients, fibre, and fruit and vegetables. And the two arms were not matched on energy density: the ultra-processed diet was the more energy dense of the two, which the authors name as a potential mechanism. The free-living design also meant energy intake and eating rate could not be measured directly Dicken 2025.

On the cardiometabolic side, the trial’s own summary is unusually candid: the greater weight and fat-mass reductions on the minimally processed diet “did not translate into significant improvements in cardiometabolic risk factors over the UPF diet, except triglycerides”. The ultra-processed diet itself reduced heart rate, fasting glucose and LDL cholesterol; total cholesterol fell on both diets; only the minimally processed diet significantly reduced blood pressure, and even that did not differ significantly from the other arm. The authors note that their secondary outcomes were exploratory and not adjusted for multiple comparisons Dicken 2025. Their conclusion is not that processing is a side issue: they write that the findings “highlight the importance of food processing in public health policy and dietary guidance in addition to existing recommendations”. They also caution that the results may not generalise to people with low ultra-processed intake, and do not imply that switching from a low- to a high-ultra-processed diet following dietary guidance is neutral or favourable Dicken 2025.

What The Cohort Evidence Adds

The cohort studies cover vastly more people than the trials do, at the cost of experimental control. In the French NutriNet-Santé cohort of 105,159 adults followed for a median of 5.2 years, with 1,409 first cardiovascular events, each absolute increment of 10 percentage points in the share of ultra-processed food in the diet was associated with a hazard ratio of 1.12 for overall cardiovascular disease (95% CI 1.05 to 1.20, P<0.001), 1.13 for coronary heart disease (1.02 to 1.24, P=0.02) and 1.11 for cerebrovascular disease (1.01 to 1.21, P=0.02). Those results held after adjustment for saturated fatty acid, sodium and sugar intakes, and separately for a healthy dietary pattern derived by factor analysis; further sensitivity analyses adjusting for dietary fibre, fruit and vegetable intake, smoking in pack-years, season and region did not change them Srour 2019. The exposure is the percentage of the diet by weight, so the estimate is per ten percentage points, not per ten per cent more. Quartile-level estimates in the same paper were frequently non-significant, and the authors are explicit that “causality remains to be established” Srour 2019.

In the Spanish SUN cohort, 8,451 middle-aged university graduates who were not overweight or obese at baseline were followed for a median of 8.9 years, yielding 1,939 incident cases of overweight or obesity. Participants in the highest quartile of ultra-processed consumption had a hazard ratio of 1.26 against those in the lowest (95% CI 1.10 to 1.45, P for trend 0.001), and the authors call for further longitudinal studies to confirm it Mendonça 2016. That study predates the 2019 guide and used its own definition of the exposure — ready-to-eat, ready-to-drink or ready-to-heat products made predominantly from components extracted or refined from whole foods — rather than the marker-ingredient rule described above, and its population is a narrow one.

An umbrella review published in The BMJ drew together 45 distinct pooled analyses from 14 published meta-analyses, with a summed total of 9,888,373 participants across those analyses — a sum in which the same people recur, not a count of distinct individuals. Direct associations appeared for 32 of the 45 health parameters examined. Four findings reached the review’s highest credibility class: cardiovascular-disease-related mortality (risk ratio 1.50, 1.37 to 1.63), type 2 diabetes on a dose-response basis (risk ratio 1.12, 1.11 to 1.13), anxiety outcomes (odds ratio 1.48, 1.37 to 1.59) and combined common mental disorder outcomes (odds ratio 1.53, 1.43 to 1.63) Lane 2024.

The quality grading runs on a different axis from credibility, and this is where the two are most often blurred together. Of the 45 pooled analyses, GRADE rated four as moderate quality, 22 as low and 19 as very low. All four moderate ratings were dose-response analyses upgraded specifically because of a dose-response gradient, and they covered all-cause mortality, prostate cancer, overweight and obesity assessed together, and type 2 diabetes. Among the four top-credibility findings, then, the type 2 diabetes dose-response association was the one that also reached moderate quality; the cardiovascular-mortality result was rated very low and the two mental-health results low Lane 2024. A further tier of associations sat one credibility class down, including all-cause mortality (risk ratio 1.21), obesity (odds ratio 1.55), depressive outcomes (hazard ratio 1.22) and adverse sleep-related outcomes (odds ratio 1.41) Lane 2024.

Thirteen of the 45 analyses fell into the review’s “no evidence” credibility class. That is a label under the review’s pre-specified classification criteria rather than a demonstrated absence of association; the same paper notes that 93 per cent of pooled analyses had point estimates in the same direction, that 24 per cent had confidence intervals including the null value, and that a low or very low rating “does not necessarily negate the potential for an association” Lane 2024. Two further points from the review are worth stating precisely, because they are routinely mangled. GRADE begins by assigning observational epidemiology a low quality rating by default, and the downgrades beyond that were attributed “largely owing to inconsistencies or heterogeneity in the effect estimates found across the original research articles or owing to imprecision” — with moderate to high heterogeneity observed across 45 per cent of the pooled analyses. On confounding, the review’s own reading is that “the consistent findings across most pooled analyses in our review support the notion that residual confounding does not fully explain the observed associations” Lane 2024.

The Published Objections To The Classification

One published critique, a single-author commentary in Current Developments in Nutrition, sets out the definitional problem in detail. Comparing the food lists used across studies, it finds “considerable variability in the lists of foods deemed to be ultra-processed” and “marked variation in the terms used to define specific ultra-processed foods”, leading to “confusion and to subjective recoding of national food-consumption databases”. Bread is its worked example, and it cuts in an unexpected direction: one French paper excluded artisanal and home-made breads from the category, while a household-budget study relied on a market report whose “artisanal bread” sales proxy itself included flour premixes and bread baked at in-store supermarket units — so supermarket in-store bread was counted as artisanal and thereby escaped the ultra-processed tally Gibney 2019.

The same critique makes a nutrient argument that is often reported backwards. Its claim is not that group four holds a wide spread of nutrient profiles; it is that the exposure fails to discriminate on several nutrients at all. Across quintiles or quartiles of ultra-processed intake, it reports, fat intakes “show little variation”, saturated fat “seem[s] not to be influenced”, and sodium “shows no tendency to rise” — declining across quintiles in Brazil. Sugars, by contrast, rise sharply and fibre falls. It also observes that “all studies to date with the NOVA classification focus on nutritional data rather than technology data”, and argues for food reformulation as a complementary lever rather than avoidance Gibney 2019. Readers should weigh it alongside the author’s declared interests, set out in the paper itself: ad hoc consultancy for Nestlé, chairing a research consortium funded by Cereal Partners Worldwide, leading a project funded by Mondelez, PepsiCo, Unilever, Nestlé and Coca-Cola, and a seat on the board of ILSI Europe Gibney 2019. The definitional criticism is acknowledged from the other side of the argument too: the umbrella review notes that “concerns [have been] raised about its possible imprecision and inconsistency” Lane 2024.

Where The Mechanism Sits

Two candidate mechanisms are worth separating. The first is behavioural. Softer food that is easier to chew and swallow may be eaten faster, delaying satiety signalling; that is the explanation the inpatient trial’s authors themselves offer for their faster eating rate and higher intake, and individual differences in eating rate between the two diets were moderately correlated with individual differences in energy intake (r=0.45, P=0.047) Hall 2019. The nine-participant 2024 crossover study that linked increased intake to reduced chewing frequency points at the same family of explanations, on a sample far too small and too narrow to settle anything Hamano 2024.

The second is biological and concerns the additives themselves. In mice given the emulsifiers carboxymethylcellulose or polysorbate 80 in drinking water at 1.0 per cent for 12 weeks, microbiota composition was altered, low-grade inflammation and features of metabolic syndrome appeared in wild-type animals, and robust colitis developed in strains already predisposed to it. That paper sets its own exposures against the regulatory position: polysorbate 80 is approved by the US Food and Drug Administration for use in select foods at up to 1.0 per cent, while carboxymethylcellulose is generally regarded as safe and used at up to 2.0 per cent but, in the authors’ words, “has not been extensively studied” Chassaing 2015. This is a mouse experiment, and the robust colitis result required genetically susceptible animals — though the authors report that supplementing chow with 1.0 per cent of either emulsifier “fully mimicked the pro-inflammatory effects and metabolic changes induced by emulsifiers in drinking water” Chassaing 2015. The authors frame the emulsifier link to inflammatory bowel disease as a hypothesis. It should be read as a mechanistic lead, not a human finding.

Neither feeding trial settles between those explanations. Both compared whole dietary patterns, so texture, eating rate, energy density, additive load and ingredient fractionation all moved together Hall 2019Dicken 2025. The inpatient authors call for future studies matching the diets more closely on protein and non-beverage energy density while including ultra-processed foods that are typically eaten slowly — a fair description of the study that would begin to separate the candidates Hall 2019.

Which Packaged Foods Carry The Nutritional Load

One finding survives every caveat above, because it is a measurement of composition rather than an estimate of risk. In a nationally representative analysis of United States dietary data from 2009–2010, ultra-processed foods supplied 57.9 per cent of total energy intake but 89.7 per cent of the energy from added sugars. Added sugars made up 21.1 per cent of the calories inside ultra-processed foods, against 2.4 per cent inside processed foods and 3.7 per cent inside minimally processed foods and culinary ingredients grouped together. Across quintiles of ultra-processed intake, added sugars rose from 7.5 per cent of energy in the lowest to 19.5 per cent in the highest, and 82.1 per cent of people in the top quintile exceeded the 10-per-cent-of-calories limit for added sugars that the paper attributes to the US Dietary Guidelines Advisory Committee, against 26.4 per cent in the bottom quintile Steele 2016. That analysis is cross-sectional, and its data are now roughly fifteen years old.

It also shows that the added-sugar burden is not spread evenly inside group four. Soft drinks alone accounted for 17.1 per cent of all added sugars in the US diet, while breads contributed 7.6 per cent of added-sugar energy at only 5.7 per cent added sugar by calories Steele 2016. The concentration is in sweetened drinks, confectionery and sweet baked goods, along with the sweetened bar-and-snack aisle we look at separately in our review of protein and snack bars. The authors’ own conclusion is that reducing ultra-processed consumption “could be an effective way of reducing the excessive intake of added sugars” Steele 2016.

Which suggests a more useful way to read a label than counting ingredients. Whether a product is technically group three or group four tells a shopper something about how it was manufactured. What the sugar, sodium and fibre lines say tells them what it will deliver nutritionally. Dietary patterns are more usually defined by what they contain than by what they exclude, and the Mediterranean pattern is one such example. The free-living trial is a reminder that the two framings need not compete: both of its arms followed the same national healthy-eating guidance, and the minimally processed version still came out modestly ahead Dicken 2025.

The people who ran these studies have said what they think follows. The inpatient authors recommend limiting ultra-processed food as a strategy for obesity prevention and treatment, one they describe as compatible with low-carb, low-fat, plant-based or animal-based approaches, while adding that policies discouraging it “should be sensitive to the time, skill, expense, and effort required to prepare meals from minimally processed foods” Hall 2019. The free-living authors call for guidance on food processing to be added to existing nutrient-based recommendations Dicken 2025. The umbrella review calls for population-level public health measures and for “urgent mechanistic research” Lane 2024. What none of those recommendations rests on is the aisle version of the rule, in which a long ingredient list is the enemy and a package is the tell. The published test is narrower than that, the effect size in free-living conditions was modest, and the added-sugar load is concentrated at the sweetened end of the category.

References